AI vision models can now keep reasoning skills while learning to see better

Researchers found that when you train vision-language AI models to get better at visual tasks, they lose the ability to reason about what they see — but a small technical fix can prevent that loss. The fix lets the model preserve internal knowledge pathways that matter for thinking, not just perceiving, which means the same model can both see accurately and reason soundly without needing separate specialized versions.

This addresses a real engineering tradeoff that has forced builders to choose between visual accuracy and reasoning ability in AI systems; if the fix generalizes, it removes that choice constraint and makes it cheaper to build models that do both.